So Intel and Nvidia dropped a CPU-GPU bomb earlier today, announcing a distinctly unexpected plan to collab' on new computing products based on Intel's x86 CPUs and Nvidia's GPUs. Oh, and there's the minor matter of Nvidia investing $5 billion in Intel. But what will these unholy Intel-Nvidia chips look like?
For now, there are virtually no official details, though we might get more this evening in the dual-CEO presser with Lip Bu Tan and Jensen Huang. But there's an existing chip that might just give some very important hints of what to expect.
I speak of the so-called GB10 "Superchip". GB10 is also a collaboration between Nvidia and another chip maker, though in this particular case, it's Mediatek and an Arm rather than an x86 CPU. However, it sports several details that immediately ring a bell in the context of the announced Intel-Nvidia deal.
In today's press release, there was a broad reference to Nvidia's high-bandwidth, low-latency NvLink interface being a critical part of the technological tie-up. Meanwhile, the consumer PC part of the deal was described as "x86 system-on-chips (SOCs) that integrate NVIDIA RTX GPU chiplets. These new x86 RTX SOCs will power a wide range of PCs that demand integration of world-class CPUs and GPUs."
Now, that GB10 chip has an SoC die from Mediatek with the CPU cores, memory controller, external interconnects like USB and PCIe, and so on. Inserted into the same package is an Nvidia GPU die. Critically, the two dies are connected via an NVLink-based 'C2C' interface. And NVLink is what the new PC project with Intel is supposed to be all about.
The Mediatek die in GB10 also has a little Nvidia IP in its display engine that sits in front of the DisplayPort and HDMI outputs. The benefit of all this is super-high bandwidth and lower latency than you get with a traditional PCI Express interface between the CPU and GPU.
You can read more about the specific benefits of NVLink in Nick's piece on GB10 here. But if you wanted to get even more speculative, you might guess that Nvidia could actually reuse the GPU die from GB10 in the first generation of Intel-Nvidia chips. The only problem with that is the assumption that Intel will need to engineer a CPU SoC with NVLink support built in, and that will take a while.
By the time that an NVLink-enabled Intel die is ready, the GPU chiplet in GB10 might be a bit long in the tooth. That said, GB10's GPU die has basically the same specs as a desktop RTX 5070 graphics card. So even in a year or two, it will offer decent graphics performance. And that's pretty exciting.
Certainly, reusing that GB10 graphics die would make it cheaper and easier for Nvidia to test the viability of this new collaboration. It would also leave Intel with all the engineering effort and cost. Nice, for Nvidia that is.
Of course, if GB10 is something of a template for these Intel-Nvidia chips, then there's at least one reason to be worried. Nvidia's DGX Spark machine that houses the GB10 chip is $4,000. Ouch.
Of course, it's unlikely that the Intel-Nvidia collaboration will result in basic PCs costing that much. But DGX Spark does hint that it might be unrealistic to expect the new Intel-Nvidia chips to enable, say, super cheap laptops with great gaming performance.
Indeed, we'll probably have to wait quite some time to get a really detailed idea about the final specs of these new chips. They could well be years away. But maybe, just maybe, the GB10 chip provides a hint at what to expect. And it could even lend one of its dies to make it all happen sooner than you might think.
